Google has unveiled Gemma, a model-based open-source language model inspired by their own Gemini commercial model and developed by the Google DeepMind team. The technical architecture of Gemma shares some similarities with Gemini.
Google has released Gemma in two sizes, 2B and 7B (available for download from Kaggle or Hugging Face), stating that the 7B model can outperform competitors like Llama 7B in various benchmarks and even surpass the larger Llama 13B model.
An interesting aspect of Gemma is that it allows full commercial use, unlike Llama, which only permits use in most cases, excluding very large companies like Meta’s competitors.
In terms of tools, Google has ensured full readiness for use on Google Cloud, with customization available for running on both Google Cloud TPU and NVIDIA GPU. Software-wise, tools are readily available on JAX, PyTorch, TensorFlow, Keras, allowing developers to design models that can run directly on their PCs.
TLDR: Google introduces Gemma, an open-source language model developed by Google DeepMind, offering commercial use and compatibility with various tools on Google Cloud and personal computers.
Leave a Comment